“Every elite school says character and the ability to overcome obstacles are priority factors in admission” How are these things measured?
Seriously? This is College Confidential, with thousands of threads and posts obsessing over all aspects of college admissions. There are seemingly infinite references to “holistic admissions” and vague exhortations to potential applicants to “read what the school is looking for” (lol). Now, the Asian discrimination documents have given us a treasure trove of raw data to mine and to infer what one of the big kahunas is really seeking, and suddenly it’s crickets, and let’s not talk about this? Odd.
For those who still care, I’ve calculated from the data I linked to in the first post that over the 10 year period covering incoming Harvard classes of 2007 through 2016, there were an average of 480 white and Asian unhooked applicants who presented a superscored 2400 on the SAT. These applicants enjoyed an admissions rate of 32.5% (white) and 26.5% (Asian), for a blended average rate of 29.1%. My guess is that there were no more than 250-350 single sitting perfect 2400 scores and that these enjoyed even higher rates of admission.
Dropping the (again) superscored SAT by just 60 points (equivalent to a 1560 on the current scale) collapses the rate to 16.5% white and just 12.0% Asian, 14.1% blended. Note that one or two “careless errors” that couldn’t be rectified on a superscore would have landed you here. There were an average of 1315 of these applicants per year.
Note that at 2340 (1560 equivalent) there were more than 2.7x the number of applicants to choose from compared with 2400 scorers, and yet the rate of admission more than than halved. I find it hard to believe that the nonacademic and other nonquantifiable attributes of this much larger group were not in aggregate greater than those of the much smaller group of perfect scorers, leading to the conclusion that Harvard found something special in this 2400 group. If anyone has an alternate explanation, please post.
At SAT 2280 (1520 equivalent) - 10.5% white rate, 6.0% Asian, 8.6% blended, 1550 applicants.
SAT 2220 (1480 equivalent) - 7.5% white, 4.5% Asian, 6.1% blended, 1400 applicants.
SAT 2160 (1440 equivalent) - 5.0% white, 4% Asian, 4.6% blended, 710 applicants.
SAT 2100 (1400 equivalent) - 4.5% white, 2.0% Asian, 3.8% blended, 585 applicants.
Below 2100 all rates go < 2%.
All numbers approximate, of course, and on average where appropriate. These are only for white and Asian applicants, excluding legacies and athletes. Presumably development and other special categories are also excluded.
I hope these data help some unhooked applicants out there to assess their chances on at least one easily observable metric that appears highly correlated with admissions. Unfortunately, there is no way in the docs I have found to get at ACT 36C scorers, although my surmise is that truly “perfect” 36C (each section) would enjoy the same sort of bump as between SAT 1560 and 1600 (equivalent).
Admissions may be more difficult today of course, but I bet not substantially so at the very highest score levels.
I’ve posted explanations earlier – correlation is not causation. A 2.7x decrease may sound like a lot, but if you look at the difference in admit rate for changes in non-academic factors, they are often much larger. Unlike academic index and presumably scores, the strong boost can still remain in the regression coefficients after full controls.
For example, in the referenced class of 2019, applicants who received a 2 in personal rating had a 12x greater admit rate than those who received a 3 personal rating. In the baseline (unhooked) sample, the regression coefficient after full controls for this boost was 2.415 (0.118) – a coefficient of 20x the standard error, so highly statistically significant after controls for the other factors, including SAT scores. Similarly applicants who received a 2 in ECs had a 6x greater admit than those who received a 3. The regression coefficient after full controls for this boost was 1.810 (0.108) – 17x the standard error, so again quite significant after controls, including SAT score.
The lawsuit and statements from the Dean of Admission suggests the vast majority of admits fall in to one of the following 3 groups:
- The rare <2% of applicants who gets a 1 in one of the core admission rating categories. Most of this group is composed of recruited athletes, but it also includes the 1 academic rating applicants. As discussed getting a 1 requires more than just perfect stats and often involves faculty review of academic work. This group had a high average admit rate of nearly 80%, but only made up 12% of admitted students.
- "All-arounders" who get high 2 ratings it at least three core categories -- usually academic, EC, and personal. This group had a 45% admit rate rate in the lawsuit, and made up 46% of admitted students. With a 45% admit rate, the majority were still rejected, so it's by no means a mechanical get three 2's and your in. Based on the Card analysis, I think it's likely the LORs, interview, and essays are important for determining which 45% among this group is accepted.. However, as noted in the first paragraph, the chance of admission drops dramatically with each 3.
- Applicants having powerful hook(s), who are are not quite at the admission levels for unhooked acceptance based on #1 and #2. For example, legacies with a 2 academic rating had a high 55% admit rate. The regression coefficients suggest legacy, suggest the boost remains after controls for the other ratings, and is a powerful enough hook to be be more relaxed about the criteria in #2.
Looking more at category #2, which makes up the largest portion of admitted students, getting three 2’s is correlated with SAT score. The strongest correlation occurs with the academic category and weakest with personal, but there is a correlation nonetheless. Note that I can predict the admit rate for unhooked white applicants (baseline) by academic decile (which is 2/3 composed of SAT scores) well by simply multiplying the rate of 2s for academic, EC, personal, and LOR, as listed below. I realize product is not valid since they are not independent variables, but it still shows that you can have a notable change in admit rate by SAT score correlations, even if you do not consider SAT scores directly in the application, beyond its contribution to being among the 42% of applicants who gets a 2 academic rating.
Rate of Academic 2 * EC 2 * Personal 2 * LOR 2 * 3.6 scaling
Top Decile – 15% (actual admit rate = 15.3%)
2nd Declie – 11% (actual admit rate = 10.8%)
3rd Declie – 7% (actual admit rate = 7.5%)
4th Decile – 5% (actual admit rate = 4.8%)
Yes, @Data10, but SAT score is observable by applicants, while the other components are not. Everyone understands correlation is not causality.
With those admit rates by score alone, we can be certain that Academic Index 10 contains many kids with less than perfect scores.
If you look at the page before the one I referenced in the initial post (page 6 of the OIR report linked) you will see admit rates of 40-50%+ at approx +1.6sd on the mechanical academic index. Some of the academic rating 1s will be there, but only a few because they are distributed throughout the top few deciles no doubt, there are very few anyway and they are not designated on the basis of scores and GPA alone. Overall, of course, admit rates in the top Academic Index decile are quite a lot less than that, suggesting a large increase in the odds for “perfect” kids within the decile. Note as well these are all unhooked kids in those data, no doubt perfect legacies approach much higher rates of admission than 50%, but there will be very few of them anyway because of restricted sample population.
What you fail to address is why the non-academic factors that Harvard is looking for should be so highly correlated with what you refer to as a small difference based on a few careless errors on a test that does not recreate Harvard’s own sorts of tests in its courses. The solution of course is really pretty simple. Any cognitively challenging task will correlate with any other. The SAT is as good as a reverse digit test would be no need to worry that the SAT is simple. Either Harvard is looking for the intellectual ability represented by perfect scores or the ability to achieve perfect scores is highly correlated with the non-academic attributes it seeks. Not every kid up there gets in of course, nor would we expect them to. Holistic does mean something. But it’s a whole lot different to tell a perfect stats kid that historically 30 to 50% were accepted than to tell them that Harvard could fill its class with perfect stats kids, which was the point of this whole thread.
BTW, I don’t care about the coefficients, the raw data is where it’s at.
The common postings of high SAT/ACT scores with not-so-high HS GPA suggest that there also exist the opposite – good test takers who do not succeed at school to the same level that they do on standardized tests.
From a college’s point of view, “unbalanced” HS GPA and SAT/ACT scores may leave questions about the applicant, though different questions depending on which one is the low one.
Of course, vocabulary was heavily coached in some high schools’ English courses (i.e. weekly vocabulary words to learn and take short quizzes on later). Even if you want the SAT to be an IQ test, it is not easy to design an IQ test that is free of environmental influences.
I’d still like to see the quote that H or any tippy top said fill the class with top scorers.
“Hard work, diligence and character do not necessarily to equate to intelligence”
And guess what? Top stats in high school don’t equate to a well done app package. All this focus on scores, because you can, still misses how it really works.
They could just administer the math and writing placement tests (which I assume are more long tail tests than the SAT/ACT) as an IQ test/entrance exam at the same time as the interview. But then you’d have something a lot more like Oxbridge which is clearly not what Harvard is aiming for.
The primary reason why the admit rates across the full academic decile as listed in the lawsuit are quite a lot lower is because the admit rates are inconsistent between the OIR report and the lawsuit analysis. The top decile should start at a standard index of 1.3, which according to the OIR has an admit rate of ~30% for white applicants. The admit rate starts at 30% and goes higher for the the upper portions of the decile. However, Arcidiacono’s lawsuit report mentions an admit rate of 15% for the full decile (baseline, white)… I don’t see an explanation for the discrepancy.
That seems unlike your usual analytical style of posting. Regression analysis can be quite useful for distinguishing between correlation and causation – Harvard looking for high scores vs high scores being correlated with other criteria Harvard looks for. When we see the regression coefficient for AI sharply drop from highly significant to less significant between model 4 and model 5, it is suggestive that the additional criteria in model 5 ( indicators for each academic, extracurricular, teacher 1, teacher 2, counselor, alumni personal, and alumni overall ratings) are more associated with with what Harvard is looking for, rather than AI score itself.
I’m sure lots of things are correlated with what Harvard values – what kind of car applicants drive, applicants favorite meals, how fast they can react to an image on a screen, and so on. Which ones are “highly correlated” is a matter of perspective. For example, the EC rating and Personal rating appear to both be more correlated with Harvard admission decisions that SAT score in isolation. And these non-academic factors are far more correlated with admission decisions than SAT when considering the full application instead of parts in isolation. The ECs and Personal rating are both only one small part of the application that does not involve direct cognitive testing, so why do ECs and Personal appear to be more highly correlated with what Harvard values than the more cognitive testing linked metrics, like SAT score?
Regardless of the correlations, Harvard has made it clear what they highly value in terms of academics. They do flag one special group, and that group gets a huge boost in chance of admission, but it’s not the perfect SAT score kids or perfect stat kids. It’s what the Harvard dean of admission describes as follows.
FWIW: I don’t believe William Fitzsimmons or any other Harvard administrator has made a statement about how they could fill the class with top test takers. However, there is a statement pertaining to valedictorians: https://theconversation.com/youre-not-going-to-get-accepted-into-a-top-university-on-merit-alone-87985
I don’t really disagree with anything you wrote there, @Data10. And that is a very helpful observation regarding the discrepancy between the Arcidiacono and OIR admit rates at the top decile, which is large (~2x). Perhaps there is something to explore in relation to development candidates, who are overwhelmingly white and who are likely to be fairly weak. In the OIR analysis, only legacy and athletes are specifically excluded, although the document is fairly informal so I am not sure it’s accurate; perhaps the addition of a large (~15% of white admits) group of development candidates in the OIR sample has the effect of shifting the mean leftwards, with corresponding shifting of the tails? As you can tell, I am just thinking out loud.
There are large hints that the stats of legacies and development are quite weak. For instance, in the Arcidiacono data in the rebuttal report (which includes all early action applicants) a relative comparison between baseline and expanded samples on the academic index is revealing. We can’t directly compare of course, because the z-scores are standardized on each sample set, but we can concentrate on the distances between a group that enjoys the least development and legacy preference (blacks) and the group that enjoys the most (whites). The z-score for black academic index rises from +0.33 in the baseline to +0.55 in the expanded - which includes the assumed weak legacies and development. Quite a jump, and a dramatic narrowing versus both whites and Asians. Another clue is the relative fall of the Asian index from iirc +0.91 to 0.82, a larger relative fall than for whites baseline to expanded. This suggests to me that the weakness may be in legacies, as development is relatively low for Asians and high for whites, which I know is contrary to my thinking out loud above. I built a quick spreadsheet that seemed to suggest that the weakness in legacy+development was largely on GPA and SAT2 measures, but it’s definitely not ready for prime time.
About regression coefficients, it’s not that I ignore them usually of course (you can tell I understand the value), it’s just that Card’s inclusion of such obviously directly manipulated measures and his non-pooled approach weaken the value of the models for me to the point where I dismiss them. There is no chance that final reader Overall rating is not goalsought, and obviously Personal Rating has been clumsily “massaged” (just look at the cross correlations with other metrics by deciles and groups). Thus, correlation of Personal Rating with logit output is to be expected, as it was intended to skew the decision. (The Overall Rating is even more outrageous - really ham-handed and an obvious attempt to try to get around Bakke, and so must be decades old I bet.) These should never be used as controls. (Steve Hsu, who is smarter than you and I put together, has made the same point.)
I tend to trust the data we can extract from the OIR analysis. Harvard never thought these would be made public, and the reports were prepared prior to the filing of the lawsuit. Again, the basic point is that I find it highly plausible that high ECs, LoRs, Recommendations, etc. will be correlated with high academic measures, including scores. But why should that be? You made the good point that 60 points on the old SAT doesn’t mean much in terms of the test’s predictive validity itself, but unless Harvard is weighting that conscientiousness to not miss any questions extremely highly, we need to make the assumption that the unobservables in that small “perfect” group are so much larger than the 2340 (1560) group as to move the admit rate from 15 to 30%. Intelligence theory provides a consistent explanation, as tasks are simply easier to accomplish for the very top of the distribution, leaving more time available for those whose behavioral and character traits allow them to achieve in other important dimensions like ECs, advanced coursework, etc.
Harvard’s flagging of Academic 1s is useful, but not dispositive as to what it is looking for in terms of academics. Clearly, they do not want too many off the chart brilliant kids around for obvious reasons relating to school cohesion, but the roughly, say, 60-70 Academic 1s who are actually admitted per year in my estimation is not enough for a school with 2000 admits. My guess is that Harvard is also looking for 200 or so +3sd kinds of kids, and they are mostly to be found in the top of that academic decile 10. After all, they do need some of their grads to get into tippy top professional schools, and +3sd will do it
I’ll sign off here on this, as CC is not supposed to be a debating society…
@Data10
About the discrepancy between the Arcidiacono and OIR reports in the top academic decile, I think we may be looking at the OIR data incorrectly. The graph is of the right side of the distribution only, starting at the mean. So, the top five academic deciles. These are only white and Asian numbers, and do not include weaker groups that would otherwise be traced on this graph, which is why I think visually it appears to cover more than 50% of a roughly normal distribution.
The deciles are not fixed in terms of academic index z-score, but rather represent roughly equal numbers of candidates ranked by index (this is consistent with Arcidiacono, roughly 2300 unhooked candidates in the top decile). This suggests the right way to transpose the OIR data into deciles is to partition the area under the two curves tracing N candidates into to five equal areas. Obviously, we cannot do that any other way than visually, but looking at it, it seems plausible that the break between decile 9 and 10 could be around +1.02sd on the chart, certainly not much higher. Slightly different constructions between the Arcidiacono deciles and the OIR no doubt fudge things up a bit (and different years) but if you squint it doesn’t seem implausible that that partitioned top decile in OIR has a weighted and blended admit rate roughly consistent with what Arcidiacono found. Also, adding in the N data from the weaker groups who do not appear on the graph would further tend to shift the break point leftwards.
If I am right, this partitioning would imply a huge spread within the top decile (~0.6sd), but intuitively that makes sense as there are so few right at the top even at Harvard. The note at the bottom of the graph would seem to explicitly confirm my intuition of very high admit rates at the very top of the academic decile 10 kids:
That 50% figure is really quite extraordinary, as nothing “holistic” goes into that index - not even the reader’s adjustment for “rigor” although perhaps it is built into the GPA-ranking formula that Harvard must use to compare across schools. Those top rated students must be the perfect scorers+4.0 GPA (who else could they be - they are up against the tail constraint).
For people following this arcana, the graph is here on page 6: http://samv91khoyt2i553a2t1s05i-wpengine.netdna-ssl.com/wp-content/uploads/2018/06/Doc-421-145-Admissions-Part-II-Report.pdf
That’s a good point. Normal distribution approximations do not work well when limited by a maximum score. Looking a the graph, I’d estimate that the top decile among the sum of Asian and White applicants starts between 1.02 and 1.12, and the average among this top decile group would be near 1.22. This is reasonably close to Arcidiacono, with assumptions about Asian/White vs full population and the varied acceptance rate in the different sample years.
The graph suggests that in the listed 2003 to 2012 period, there was an average of under 50 non legacy/URM applicants per year who had the maximum listed academic index, which would correspond to being valedictorian in a HS class size of 300+ (class rank was typically used in AI during this period, and most applicants submitted rank), perfect SAT, and perfect SAT II. The earlier SAT graphs and lawsuit numbers suggest far larger numbers of perfect SAT/ACT score applicants than 50 per year, so the vast majority of them were likely not meeting the class rank (or sometimes GPA) criteria.
Agreed for the most part, but as for shape of the curves here the maximum score is only one part of the index, and is really only a limit for one demographic (Asian). You have to imagine all the other distributions added in, including the weaker hooked groups, because the deciles are calculated on more data then shown. Visually you can see that the nonhooked Asians and whites exhibit negative skew (left tailed), with Asian especially truncated, while URM curves (not shown) are all going to be positive skew/long-tailed right, with I bet the other categories roughly bell shaped. There are obviously correlations among the components in the academic index (high between SAT and SAT2, moderate between SAT and GPA), but when aggregating there is probably enough to ensure a rough bell shape to AI when all curves are thrown into the mix (technically central limit theorem isn’t applicable due to correlations, but it sort of is :)). Not normal, but “roughly” close.
I especially agree that GPA/rank is likely to be the limiting factor, and I can’t tell how Harvard calculates it. Nevertheless, it is probably similar to how it was described in Michelle Hernandez’s book a generation ago - these guys don’t change too much as a general rule, except in response to lawsuits or SCOTUS decisions - and would put kids from smaller schools at a disadvantage. 2400 SAT and perfect SAT2 in the age of superscoring does not seem too tough to me; 800s on math and sciences {and perhaps languages for Asians) was pretty easy imo especially as only the top scores would have been submitted and subject choice was applicant’s, the tough part was likely the required essay scored as part of the writing section.
I do think we have made a lot of progress. “Perfect stat” kids do in fact see tremendous advantage at Harvard, with the notable hurdle being GPA/rank. 30-50% admit rates historically on just a stats only measure is pretty compelling imo.
It might be time to start a new thread to try to untangle the legacy and development stats, extracting them from the Arcidiacono appendices. Another often thrown around theme on here is that legacies and even development admits are likely to be just as strong academically as the nonhooked admits. That’s of course highly improbable, and the z-score data hint at it. With some work, I bet we can figure out just how weak they are.
I throw this in only as a broad piece of information that may illuminate an area where the data is not provided --almost as a back door to understanding where different applicants fall in academic scoring.
RE athletes: They are the only ones who have to reach a minimum academic level, at least on an overall team basis, generally within one standard deviation of the average Academic Index of the whole school. It’s a given that there are many students who have higher grades and scores than athletes. Those are the ones who have above average AI’s (all students have an AI score with a max of 240–so a 240 AI would be a “1”). If one assumes that the average is pretty close to the median, though, then there are just as many students who are below the average—and below a lot of the athletes. So its ironic that people question the admissibility of athletes–that they would be questioned in the framework of the lawsuit when the athletes are the only group held to a minimum. The minimum is 176 and the team average is about 218/220 depending on the overall student body AI. So the more the perfect scores and grades there are–the higher the mean will be as well as the converse. What would be really interesting is to see AI mean by category of admit.
Non-recruits don’t get the bye recruits can get. The team needs to meet the AI, not individuals.
You can dissect it any way you want but there aren’t that many academic superstars available to fill up ivy league schools. Even among small pool of these superstars, not everyone want to come to Ivies or can afford Ivies, many head towards MIT, Cal tech, Rice, U Chicago, Berkeley, CMU, JHU while others prefer Amherst, Williams, Pomona, Harvey Mudd and military academies.
When ivies reject students of this caliber, it’s almost always because their aren’t that many seats for unhooked merit category. Admissions are not fair and transparency would expose it.
“at least on an overall team basis” and I wasn’t talking about non recruits.
Non-recruits are already in all these data we have been discussing, When Harvard says it is excluding athletes and legacies from those graphs, they mean Athletic Rating 1 recruits (which have a greater than 80% acceptance rate), and presumably not the 2 and 2+ athletic rated non-recruits.
There are plenty of academic superstars applying to Ivies. Day after day of them, top scores, no B grades. Or maybe just one, in some non-core.
There are not that many academic superstars who knock their actual apps out of the park. Big difference. These schools are not just looking for stats and some club titles or research work. The problem is not “unhooked.” It’s the apps, supps, LoRs.
The fact these colleges are holistic, not rack and stack, is completely clear.